Detect defects your team has never seen before. Our AI trains on good parts only and flags any deviation, including defects that have never occurred in production before.
Most inspection systems need defect examples to learn from. Anomaly detection works the other way around. The AI trains exclusively on good parts, builds a precise model of what normal looks like, and detects anything that deviates. That includes defects that have never been seen before. Start from a handful of reference images.
GLAD is the AI model developed by Visionairy in partnership with CNRS and ENS Paris-Saclay. It trains exclusively on compliant parts. A small set of reference images is enough to get started, with no defect examples needed. GLAD detects known and unknown anomalies simultaneously, analysing parts at global and local level, at a precision that ranks among the highest on industrial benchmarks.
When defects are too rare to collect enough examples for supervised training. The AI learns from good parts only, so no defect catalog is required.
When you need inspection running from day one, without waiting to accumulate defect history. A few reference images and you are ready.
When the defect taxonomy is not established or evolves over time. The AI flags anything abnormal, and your team classifies it afterwards.
Machined, moulded, or formed parts across automotive, electronics, and plastics.
Shape, surface, and contamination anomalies on organic products at line speed.
Micro-defects, inclusions, and coating anomalies on lenses and precision glass.
Seal quality, deformation, and surface anomalies on packaging before dispatch.
Foreign objects and contaminants in recycling or food production streams.
Weave defects, stains, and surface irregularities on continuous materials.
No need for thousands of labeled images. A small set of compliant parts is enough to train a production-ready model.
Because the AI learns normality, it catches unknown defects, not just the ones in your training catalog.
When product references change or production conditions shift, the model retrains from new examples in minutes.
Our vision engineers analyze your production case and deliver a tailored feasibility report at no cost, in under five days.
